Artificial Intelligence
& Energy
Bertrand Cornélusse,
Raphaël Fonteneau
The Smartgrids team is part of the Montefiore
Research Unit of the ULg, contains around 15
researchers and is headed by Pr. Damien Ernst
Our vision of Artificial Intelligence
Roadmap
Artificial
intelligence
Optimization
Machine
learning
Reinforcement
learning
4Roadmap
Machine
learning
Machine learning is about extracting {patterns,
knowledge, information} from data
Cluster images SIRI Cortona OK Google Convert voice signal into sentences Make on-line recommandations Recognize patterns in images Interpret sentences Google photos 6Machine learning studies and builds algorithms
that learn from and make predictions on data
Supervised Learning in a nutshell:
Imagine you have a set of data
{(x1, y1), (x2, y2), …, (xn, yn)}
represented by black points on the figure.
To be able to estimate the value of an output y for any input x, You “train” a Machine Learning algorithm using these data. You obtain the blue line.
The quality of the estimate depends on data quality/quantity: with more points, e.g. the black circles, you would for instance get the red curve.
7
x y
Recent advances in machine learning
Machine learning algorithms have recently shown impressive results, in particular when input data are images: this has led to the identification of a subfield of Machine Learning called Deep Learning.
The term “deep” refers to the fact that those learning architectures, mainly
Artificial Neural Networks, are made of several layers.
Zoom on a neuron
Deep neural network architectures
Source: http://www.ais.uni-bonn.de/deep_learning/images/Convolutional_NN.jpg
Wait… ANN are not new, right?
ANN date back to the sixties. Training ANN was not an easy task until recently. Recent progress is twofold: • Smart(er) training approaches • GPU calculus 10Optimization
Artificial
intelligence
Roadmap
Machine
learning
Reinforcement
learning
11From supervised learning to reinforcement learning
Supervised learning techniques (in particular
(deep) convolutional networks) may be used as a block in a more complex structure, in particular in Dynamic Programming (DP) or Model Predictive Control (MPC) schemes.
This connects to reinforcement learning, an area of machine learning originally inspired by behaviorist psychology, concerned with how software agents ought to take actions in an environment so as to maximize some notion of cumulative reward.
Deep reinforcement learning combines deep
learning with reinforcement learning (and, consequently, in DP / MPC schemes). 12 Agent Environment Action Reward
Playing Atari with deep reinforcement learning
13
At Google Deepmind At ULg
Human-level control through deep reinforcement learning. Nature, 2015.
Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A. Rusu, Joel Veness, Marc G. Bellemare, Alex Graves, Martin Riedmiller, Andreas K. Fidjeland, Georg Ostrovski, Stig Petersen Charles Beattie, Amir Sadik, Ioannis Antonoglou, Helen King, Dharshan Kumaran, Daan Wierstra, Shane Legg & Demis Hassabis
Breaking news
Recent breakthroughs in the field of AI for the game of GO have been done by Google Deepmind.
These results have been obtained by combining Deep Convolutional Networks with Monte Carlo Tree Search techniques. The resulting agent, AlphaGo, achieved 99.8% winning rate against other GO AI, and defeated the European Go champion by 5 games to 0. 14 Mastering the game of Go with deep neural networks and tree search. Nature, 2016. David Silver, Aja Huang, Chris J. Maddison, Arthur Guez, Laurent Sifre, George van den Driessche, Julian Schrittwieser, Ioannis Antonoglou, Veda Panneershelvam, Marc Lanctot, Sander Dieleman, Dominik Grewe, John Nham, Nal Kalchbrenner, Ilya Sutskever, Timothy Lillicrap, Madeleine Leach, Koray Kavukcuoglu, Thore Graepel & Demis Hassabis
Want to know more?
Google is launching a new deep learning course (in collaboration with Udacity): https://www.udacity.com/course/deep-learning--ud730 You may also be interested in NVidia Deep Learning course: https://developer.nvidia.com/deep-learning-courses Or even Stanford Mooc about Machine Learning: https://www.coursera.org/learn/machine-learning 15Optimization
Artificial
intelligence
Roadmap
Machine
learning
Reinforcement
learning
16Optimization: decide the values that some variables can take,
under a set of contraints, so as to maximize an objective.
A long tradition of numerical solutions and theoretical analysis.
Given assumptions on models, one can eventually get guarantees
about solutions.
How is optimization connected to machine learning?
Learning problems can be casted as optimization problems
How is machine learning connected to optimization?
Machine learning actually solves some (or part of) optimization
problems (e.g: RL, or tuning of an algo, or proxy to an algo)
Machine learning is tightly coupled to optimization
Machine learning is tightly coupled to optimization
An illustration of the simplex
algorithm. The simplex
algorithm was invented by G.
Dantzig. It dates back to the
second world war.
This can be used to solve
many practical optimization
problems.
set NUTRIENT ordered; set FOOD ordered;
param cost {FOOD} >= 0;
param minNutrient {NUTRIENT} >= 0;
param maxNutrient {i in NUTRIENT} >= minNutrient[i]; param amount {NUTRIENT,FOOD} >= 0;
# Variables
var Buy {j in FOOD} integer;
# Objective
minimize Total_Cost: sum {j in FOOD} cost[j] * Buy[j];
(or minimize nutrient_amount {i in NUTRIENT}: sum {j in FOOD} amount[i,j] * Buy[j];) # Constraints
subject to Diet {i in NUTRIENT}:
minNutrient[i] <= sum {j in FOOD} amount[i,j] * Buy[j] <= maxNutrient[i];
Optimization relies on an analytical model ...
Example: Building the lunch menu, a first application of AI for energy ;)
set NUTRIENT ordered; set FOOD ordered; param cost {FOOD} >= 0;
param minNutrient {NUTRIENT} >= 0;
param maxNutrient {i in NUTRIENT} >= minNutrient[i]; param amount {NUTRIENT,FOOD} >= 0;
# Variables
var Buy {j in FOOD} integer;
# Objective
minimize Total_Cost: sum {j in FOOD} cost[j] * Buy[j];
(or minimize nutrient_amount {i in NUTRIENT}: sum {j in FOOD} amount[i,j] * Buy[j];)
# Constraints
subject to Diet {i in NUTRIENT}:
minNutrient[i] <= sum {j in FOOD} amount[i,j] * Buy[j] <= maxNutrient[i];
Optimization relies on an analytical model ...
Example: Building the lunch menu, a first application of AI for energy ;)
+ Data
Your lunch menu
Optimization relies on an analytical model, machine learning may
not
+ + + 21Optimization and Machine learning have different aims
In the optimization world, a method targets one problem class,
or even an instance of a problem, and a theory is obsessed by
optimality (can I prove it mathematically?) and efficiency (can I
compute it efficiently?)
Machine learning is focused on statistical significance (reaching
a trade off between overfitting and “misrepresentation”),
replicability to other problems with few adaptation, and
interpretability of results
Roadmap
Artificial
intelligence
Optimization
Machine
learning
Reinforcement
learning
23Energetic applications
World energy consumption outlook around 2010
Fossil fuels Nuclear Renewables
Biomass heat Solar hotwater Geothermal heat Hydropower Ethanol Biodiesel Biomass electricity Wind power Geothermal electricity Solar PV power Solar CSP Ocean power Sources: IEA 25
Optimization has plenty of applications in the Energy industry
Electrical power systems:
• Production planning: unit commitment
• Managing grid constraints: optimal power flow
Oil and gas industry:
• Where to dig? In which sequence?
Logistics and transportation:
• Vehicle Routing Problems
Industrial processes:
• Reduction or displacement of energy consumption
26Example: Day-ahead electricity prices in
Europe are determined by Euphemia
EUPHEMIA is the market coupling algorithm for European Power exchanges, implemented and developed in-house by N-SIDE, a spin-off of UCL and ULg
Used daily by Power Exchanges to fix pan-EU day-ahead electricity prices in 19 EU countries.
Computing market prices & volumes by: • coupling national markets
• maximizing total economical welfare • optimizing network capacity utilization
• modeling complex economical constraints
Extension to whole Europe in progress
http://energy.n-side.com/day-ahead/
Evolution of the energy system
Global Grid(s) versus Microgrids
Prof. Damien Ernst - University of Liège
ELIA Stakeholders days
From decentralization…
From decentralization to centralization
From decentralization to centralization, and back
Why are we now talking about AI,
and not just about optimization?
We are now trying to optimize more and more locally, because
renewable energy sources are distributed, data is ubiquitous
and computation power as well.
However, the ratio “gain / (time to spend for gathering the data
and solving the problem)” is way smaller than for large
centralized projects.
AI offers the possibility to automate the data gathering,
modeling and optimization stages. For instance, learn from the
habits of users of a house, propose some car pooling options,
correlate all this with calendar events.
Rethinking the operation of distribution systems
Active network management.Smart modulation of generation sources, loads and storages so as to operate safely the electrical network without having to rely on significant investments in infrastructure.
GREDOR project.
Redesigning in an integrated way the whole decision chain used for managing distribution networks in order to perform active network management optimally (i.e., maximisation of social welfare).
www.gredor.be
Empowering consumers and distributed generation
Microgrids are modern, localized, small-scale
grids, contrary to the traditional, centralized electricity grid (macrogrid).
Some microgrids can operate disconnected from the centralized grid and operate autonomously, strengthen grid resilience and help mitigate grid disturbances.
Optimizing the sizing and the operation of a microgrid requires both optimization and AI techniques.
35